Summary

  • The only Lambda commitment attached to a clock is an initial support-response table: one hour for Severity 1, two for Severity 2, six for Severity 3 and 24 hours for service requests — a deadline for a human reply, not for the GPUs coming back.
  • The Cloud Terms supply the service as-is, award credits at Lambda's discretion, and cap liability at twelve months of fees, so the reported 99.99% stress-test figure carries no customer measurement period and no remedy attached.

The invoice is the easy part. On-demand instances are priced hourly and metered in one-minute increments that start after launch health checks and stop at termination, and an instance that keeps running while doing nothing keeps billing; refunds arrive as service credits toward future usage, are not redeemable for cash, and expire after twelve months (billing documentation). Everything around that meter is thinner than the catalogue suggests.

Begin with what Lambda publishes about the hardware itself. The 1-Click Clusters introduction lists 400 Gb/s Quantum-2 InfiniBand, peer-to-peer GPUDirect RDMA at up to 3200 Gb/s and two 100 Gb/s Ethernet network cards on each node (cluster documentation). Those are specific, checkable statements about equipment. They are also configuration. The same page says Direct Internet Access is delivered over shared upstream capacity at the datacenter edge and is not dedicated per tenant, which turns the Ethernet figure from a promise about your traffic into a description of the port on the server.

The on-demand table is similarly precise and similarly scoped. It carries a December 2025 label, and its eight-GPU B200 configuration lists 180 GB of memory per GPU, 208 vCPUs, 2900 GiB of RAM and a 22 TiB root volume (on-demand overview). That page declines to promise any instance type in every region, and the API exposes a regions_with_capacity_available field so a scheduler can ask where capacity is instead of assuming it (API documentation). The API reference is careful in a way worth noticing: it limits clients to one request per second and one instance launch per twelve seconds, and it lists upstream network outages, service unavailability and resource exhaustion among the errors a caller should expect to handle.

Then comes the one number that runs on a clock. Lambda's support page publishes an initial-response table for 1-Click Clusters: one hour for Severity 1, two hours for Severity 2, six hours for Severity 3, twenty-four hours for service requests (support service levels). Severity 1 is defined to include services being down, inaccessible or unable to perform useful work, so total failure is precisely the condition that starts the shortest timer in the document. The timer ends when Lambda responds. It does not end when the GPUs come back. That page describes 24/7 ticketing support and supplies no uptime percentage, no downtime calculation, no service-credit table and no compensation formula for a response that misses its hour.

Unavailability is covered elsewhere in the framework, in the general rather than the specific. The Cloud Terms supply services as-is, subject to express provisions and applicable law (Cloud Terms); award credits at Lambda's discretion, revocable and non-cash, generally expiring after twelve months; and cap liability at the amounts paid and due in the preceding twelve months, with consequential damages excluded. Termination for material breach runs through notice and a thirty-day cure, or an acceptable cure plan within ten days where a cure takes longer. Two clauses do more work than the rest: Orders specify service terms and fees and override conflicting provisions, and the agreement supersedes earlier marketing materials. The published documents are the floor; the Order is where the ceiling gets set.

One page shows that ceiling exists without showing its height. The Managed Slurm documentation describes continuous health checking, isolation of unhealthy nodes and escalation to hands-on repair, then says repairs or replacements follow the terms of your service level agreement (auto-remediation). It does not reproduce that agreement, quote it or link it. The public contract layer is therefore not the whole contract, and nothing here establishes that a buyer cannot negotiate availability terms. It establishes that Lambda does not publish them.

The most quotable uptime number in Lambda's own material is a test result. An MLPerf article states that the infrastructure maintained 99.99% uptime during extensive stress tests (MLPerf write-up), on a page displaying an April 2025 publication date and an October 2026 update. The percentage comes with no customer measurement window, no exclusion list, no claims procedure and no remedy. It is a report about a test, which is a different instrument from a service level.

The operational record has the same shape. At retrieval the status page listed the Cloud API, network, virtual machines, storage and power as fully operational, and published no target, no calculation method, no exclusions and no compensation schedule (status page).

Price, meanwhile, is a moving document. The instance prices currently displayed take effect on 16 November 2026 at 00:00 UTC, after the date of this piece, and the page directs readers to their account for pricing before then (pricing). The same page advertises 1-Click Clusters from 16 to 2,000+ GPUs while the cluster introduction describes 16 to 512, a gap the publisher does not reconcile, and it describes self-service instance access as first-come.

Put the pieces in order and the picture is not sinister, just narrow. A buyer can compute the cost of a run to the cent, confirm the fabric that carries it, and read how quickly somebody will answer when it breaks. What the buyer cannot read is a number that obliges Lambda to keep eight accelerators serving traffic. The only remedy with a schedule attached to it is a reply.

The consequence follows from the clause that gives the Order precedence. If an availability target, a restoration deadline or a credit percentage matters to a training programme, it has to be written into the Order, because the public layer will not supply it and the marketing layer is expressly superseded. The catalogue tells you what the hardware is. Only the contract can tell you what the service is. The entity record behind this analysis sits in the BTW directory.